Mu Qiao

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Claudio S. Pinhanez, German H. Flores, Marisa A. Vasconcelos et al. · ibm-research

How can we best address the dangerous impact that deep learning-generated fake audios, photographs, and videos (a.k.a. deepfakes) may have in personal and societal life? We foresee that the availability of cheap deepfake technology will create a second wave of disinformation where people will receive specific, personalized disinformation through different channels, making the current approaches to fight disinformation obsolete. We argue that fake media has to be seen as an upcoming cybersecurity problem, and we have to shift from combating its spread to a prevention and cure framework where users have available ways to verify, challenge, and argue against the veracity of each piece of media they are exposed to. To create the technologies behind this framework, we propose that a new Science of Disinformation is needed, one which creates a theoretical framework both for the processes of communication and consumption of false content. Key scientific and technological challenges facing this research agenda are listed and discussed in the light of state-of-art technologies for fake media generation and detection, argument finding and construction, and how to effectively engage users in the prevention and cure processes.

Qiying Yu, Zheng Zhang, Ruofei Zhu et al. · tsinghua

Inference scaling empowers LLMs with unprecedented reasoning ability, with reinforcement learning as the core technique to elicit complex reasoning. However, key technical details of state-of-the-art reasoning LLMs are concealed (such as in OpenAI o1 blog and DeepSeek R1 technical report), thus the community still struggles to reproduce their RL training results. We propose the $\textbf{D}$ecoupled Clip and $\textbf{D}$ynamic s$\textbf{A}$mpling $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{DAPO}$) algorithm, and fully open-source a state-of-the-art large-scale RL system that achieves 50 points on AIME 2024 using Qwen2.5-32B base model. Unlike previous works that withhold training details, we introduce four key techniques of our algorithm that make large-scale LLM RL a success. In addition, we open-source our training code, which is built on the verl framework, along with a carefully curated and processed dataset. These components of our open-source system enhance reproducibility and support future research in large-scale LLM RL.

Wenqi Wei, Mu Qiao, Divyesh Jadav

Graph Neural Networks (GNNs) have enjoyed wide spread applications in graph-structured data. However, existing graph based applications commonly lack annotated data. GNNs are required to learn latent patterns from a limited amount of training data to perform inferences on a vast amount of test data. The increased complexity of GNNs, as well as a single point of model parameter initialization, usually lead to overfitting and sub-optimal performance. In addition, it is known that GNNs are vulnerable to adversarial attacks. In this paper, we push one step forward on the ensemble learning of GNNs with improved accuracy, generalization, and adversarial robustness. Following the principles of stochastic modeling, we propose a new method called GNN-Ensemble to construct an ensemble of random decision graph neural networks whose capacity can be arbitrarily expanded for improvement in performance. The essence of the method is to build multiple GNNs in randomly selected substructures in the topological space and subfeatures in the feature space, and then combine them for final decision making. These GNNs in different substructure and subfeature spaces generalize their classification in complementary ways. Consequently, their combined classification performance can be improved and overfitting on the training data can be effectively reduced. In the meantime, we show that GNN-Ensemble can significantly improve the adversarial robustness against attacks on GNNs.

Siqi Fan, Yuguang Xie, Bowen Cai et al.

Understanding the chemical structure from a graphical representation of a molecule is a challenging image caption task that would greatly benefit molecule-centric scientific discovery. Variations in molecular images and caption subtasks pose a significant challenge in both image representation learning and task modeling. Yet, existing methods only focus on a specific caption task that translates a molecular image into its graph structure, i.e., OCSR. In this paper, we propose the Optical Chemical Structure Understanding (OCSU) task, which extends low-level recognition to multilevel understanding and aims to translate chemical structure diagrams into readable strings for both machine and chemist. To facilitate the development of OCSU technology, we explore both OCSR-based and OCSR-free paradigms. We propose DoubleCheck to enhance OCSR performance via attentive feature enhancement for local ambiguous atoms. It can be cascaded with existing SMILES-based molecule understanding methods to achieve OCSU. Meanwhile, Mol-VL is a vision-language model end-to-end optimized for OCSU. We also construct Vis-CheBI20, the first large-scale OCSU dataset. Through comprehensive experiments, we demonstrate the proposed approaches excel at providing chemist-readable caption for chemical structure diagrams, which provide solid baselines for further research. Our code, model, and data are open-sourced at https://github.com/PharMolix/OCSU.

Ruiying Lu, Bo Chen, Guanliang Liu et al.

Dual-view snapshot compressive imaging (SCI) aims to capture videos from two field-of-views (FoVs) using a 2D sensor (detector) in a single snapshot, achieving joint FoV and temporal compressive sensing, and thus enjoying the advantages of low-bandwidth, low-power, and low-cost. However, it is challenging for existing model-based decoding algorithms to reconstruct each individual scene, which usually require exhaustive parameter tuning with extremely long running time for large scale data. In this paper, we propose an optical flow-aided recurrent neural network for dual video SCI systems, which provides high-quality decoding in seconds. Firstly, we develop a diversity amplification method to enlarge the differences between scenes of two FoVs, and design a deep convolutional neural network with dual branches to separate different scenes from the single measurement. Secondly, we integrate the bidirectional optical flow extracted from adjacent frames with the recurrent neural network to jointly reconstruct each video in a sequential manner. Extensive results on both simulation and real data demonstrate the superior performance of our proposed model in a short inference time. The code and data are available at https://github.com/RuiyingLu/OFaNet-for-Dual-view-SCI.

Yuan Gao, Mu Qiao

The growth in the use of online advertising to foster brand awareness over recent years is largely attributable to the ubiquity of social media. One pivotal technology contributing to the success of online brand advertising is frequency capping, a mechanism that enables marketers to control the number of times an ad is shown to a specific user. However, the very foundation of this technology is being scrutinized as the industry gravitates towards advertising solutions that prioritize user privacy. This paper delves into the issue of reach measurement and optimization within the context of $k$-anonymity, a privacy-preserving model gaining traction across major online advertising platforms. We outline how to report reach within this new privacy landscape and demonstrate how probabilistic discounting, a probabilistic adaptation of traditional frequency capping, can be employed to optimize campaign performance. Experiments are performed to assess the trade-off between user privacy and the efficacy of online brand advertising. Notably, we discern a significant dip in performance as long as privacy is introduced, yet this comes with a limited additional cost for advertising platforms to offer their users more privacy.

Mu Qiao

Transformer architectures have revolutionized artificial intelligence (AI) through their attention mechanisms, yet the computational principles underlying their success remain opaque. We present a novel theoretical framework that reinterprets the core computation of attention as Pavlovian conditioning. Our model finds a direct mathematical analogue in linear attention, which simplifies the analysis of the underlying associative process. We demonstrate that attention's queries, keys, and values can be mapped to the three elements of classical conditioning: test stimuli that probe associations, conditional stimuli (CS) that serve as retrieval cues, and unconditional stimuli (US) that contain response information. Through this lens, we suggest that each attention operation constructs a transient associative memory via a Hebbian rule, where CS-US pairs form dynamic associations that test stimuli can later retrieve. Our framework yields several theoretical insights grounded in this linearized model: (1) a capacity theorem showing that attention heads can store O($\sqrt{d_k}$) associations before interference degrades retrieval; (2) an error propagation analysis revealing fundamental architectural trade-offs of balancing model depth, width, and head redundancy to maintain reliability; and (3) an understanding of how biologically plausible learning rules could enhance transformer architectures. By establishing this deep connection, we suggest that the success of modern AI may stem not from architectural novelty alone, but from implementing computational principles that biology optimized over millions of years of evolution.

Mu Qiao

Identifying evolutionary correspondences between cell types across species is a fundamental challenge in comparative genomics and evolutionary biology. Existing approaches often rely on either reference-based matching, which imposes asymmetry by designating one species as the reference, or projection-based matching, which may increase computational complexity and obscure biological interpretability at the cell-type level. Here, we present OT-MESH, an unsupervised computational framework leveraging entropy-regularized optimal transport (OT) to systematically determine cross-species cell type homologies. Our method uniquely integrates the Minimize Entropy of Sinkhorn (MESH) technique to refine the OT plan, transforming diffuse transport matrices into sparse, interpretable correspondences. Through systematic evaluation on synthetic datasets, we demonstrate that OT-MESH achieves near-optimal matching accuracy with computational efficiency, while maintaining remarkable robustness to noise. Compared to other OT-based methods like RefCM, OT-MESH provides speedup while achieving comparable accuracy. Applied to retinal bipolar cells (BCs) and retinal ganglion cells (RGCs) from mouse and macaque, OT-MESH accurately recovers known evolutionary relationships and uncovers novel correspondences, one of which was independently validated experimentally. Thus, our framework offers a principled, scalable, and interpretable solution for evolutionary cell type mapping, facilitating deeper insights into cellular specialization and conservation across species.

Mu Qiao

Navigating the complex landscape of single-cell transcriptomic data presents significant challenges. Central to this challenge is the identification of a meaningful representation of high-dimensional gene expression patterns that sheds light on the structural and functional properties of cell types. Pursuing model interpretability and computational simplicity, we often look for a linear transformation of the original data that aligns with key phenotypic features of cells. In response to this need, we introduce factorized linear discriminant analysis (FLDA), a novel method for linear dimensionality reduction. The crux of FLDA lies in identifying a linear function of gene expression levels that is highly correlated with one phenotypic feature while minimizing the influence of others. To augment this method, we integrate it with a sparsity-based regularization algorithm. This integration is crucial as it selects a subset of genes pivotal to a specific phenotypic feature or a combination thereof. To illustrate the effectiveness of FLDA, we apply it to transcriptomic datasets from neurons in the Drosophila optic lobe. We demonstrate that FLDA not only captures the inherent structural patterns aligned with phenotypic features but also uncovers key genes associated with each phenotype.

Yuya Jeremy Ong, Mu Qiao, Divyesh Jadav

Multivariate time series prediction has applications in a wide variety of domains and is considered to be a very challenging task, especially when the variables have correlations and exhibit complex temporal patterns, such as seasonality and trend. Many existing methods suffer from strong statistical assumptions, numerical issues with high dimensionality, manual feature engineering efforts, and scalability. In this work, we present a novel deep learning architecture, known as Temporal Tensor Transformation Network, which transforms the original multivariate time series into a higher order of tensor through the proposed Temporal-Slicing Stack Transformation. This yields a new representation of the original multivariate time series, which enables the convolution kernel to extract complex and non-linear features as well as variable interactional signals from a relatively large temporal region. Experimental results show that Temporal Tensor Transformation Network outperforms several state-of-the-art methods on window-based predictions across various tasks. The proposed architecture also demonstrates robust prediction performance through an extensive sensitivity analysis.

Mu Qiao, Luis Bathen, Simon-Pierre Génot et al.

Hybrid cloud is an integrated cloud computing environment utilizing a mix of public cloud, private cloud, and on-premise traditional IT infrastructures. Workload awareness, defined as a detailed full range understanding of each individual workload, is essential in implementing the hybrid cloud. While it is critical to perform an accurate analysis to determine which workloads are appropriate for on-premise deployment versus which workloads can be migrated to a cloud off-premise, the assessment is mainly performed by rule or policy based approaches. In this paper, we introduce StackInsights, a novel cognitive system to automatically analyze and predict the cloud readiness of workloads for an enterprise. Our system harnesses the critical metrics across the entire stack: 1) infrastructure metrics, 2) data relevance metrics, and 3) application taxonomy, to identify workloads that have characteristics of a) low sensitivity with respect to business security, criticality and compliance, and b) low response time requirements and access patterns. Since the capture of the data relevance metrics involves an intrusive and in-depth scanning of the content of storage objects, a machine learning model is applied to perform the business relevance classification by learning from the meta level metrics harnessed across stack. In contrast to traditional methods, StackInsights significantly reduces the total time for hybrid cloud readiness assessment by orders of magnitude.

Mu Qiao, Jia Li

We investigate a Gaussian mixture model (GMM) with component means constrained in a pre-selected subspace. Applications to classification and clustering are explored. An EM-type estimation algorithm is derived. We prove that the subspace containing the component means of a GMM with a common covariance matrix also contains the modes of the density and the class means. This motivates us to find a subspace by applying weighted principal component analysis to the modes of a kernel density and the class means. To circumvent the difficulty of deciding the kernel bandwidth, we acquire multiple subspaces from the kernel densities based on a sequence of bandwidths. The GMM constrained by each subspace is estimated; and the model yielding the maximum likelihood is chosen. A dimension reduction property is proved in the sense of being informative for classification or clustering. Experiments on real and simulated data sets are conducted to examine several ways of determining the subspace and to compare with the reduced rank mixture discriminant analysis (MDA). Our new method with the simple technique of spanning the subspace only by class means often outperforms the reduced rank MDA when the subspace dimension is very low, making it particularly appealing for visualization.